Article | Smart Relativity

At the main industrial development forum of the 2024 World Artificial Intelligence Conference and the High-Level Meeting on Global Governance of Artificial Intelligence, Robin Li, the founder, chairman, and CEO of Baidu, shared his views on AI large models, shocking the audience.

He first pointed out that "the Hundred Model War has caused a huge waste of social resources, especially computational power. But at the same time, it has also enabled us to establish the ability to catch up with the world's most advanced base models."

He then emphasized, "Without applications, having only base models, whether open-source or closed-source, is worthless." At the same time, Robin Li also stated that we should break out of the logic of the mobile era and avoid falling into the "super app trap," not only applications with 1 billion DAU are considered successful.

It can be said that Robin Li's speech was quite intense. This seems to be the first time that a leader has openly discussed the "Hundred Model War" and the development of large models at such a high-level occasion.

Of course, Robin Li is not the only one who holds a similar view.

Zhu Xiaohu, the managing partner of GSR Ventures, also mentioned in a June lecture that many entrepreneurs are blindly investing in AI underlying technologies. While this has created the grandeur of the "Hundred Model War," it has also caused a waste of social resources.

He emphasized, "It is evident that the direction of AI entrepreneurship has completely changed."

How has it changed?

Besides Robin Li and Zhu Xiaohu, leaders such as Fu Sheng, the chairman and CEO of Cheetah Mobile and the chairman of Orion Star, Zhang Fan, the COO of Zhipu AI, and Wang Xiaochuan, the founder of Baichuan Intelligence, have also discussed the direction of competition shifts in the large model industry on various occasions. The key point that ultimately led to consensus was "scenarios" and "applications".

The "Hundred Model War" caused by the focus on base large models seems to have come to an end. The focus of large models should now shift to "rolling up" scenario applications.

Industry leaders have reached a consensus on this point. Everyone's tone has changed this year!

Don't over-compete on base large models, it's time to "roll up" scenario applications

In the past period, the US has seen a influx of startup companies focused on large model application development, such as Adept, Stability.ai, Runway, BettrData, Tinybird, UnSkript, and many others.

Meanwhile, leading large model companies like OpenAI and Anthropic, as well as technology giants like Google and Microsoft, are also committed to developing solutions for various application scenarios using open-source models or independently developed base models.

The launch of GPTs and OpenAI's series of announcements aimed at benefiting developers are intended to attract more entrepreneurial teams to participate in the innovation and application of GPT technology, thereby enriching the GPT ecosystem and helping OpenAI gain an advantage in scenario applications in the large model field in the future.

Judging from the trends in the international large model industry, the shift in tone among domestic leaders is not unfounded.

Currently, the daily average invocation volume of Baidu's Wenxin Yiyan has exceeded 500 million, while just two months ago, Baidu officially announced that the daily invocation volume of Wenxin Yiyan had exceeded 200 million.

The significant change in invocation volume over a period of two months shows that the focus of large models on scenario applications is not only driven by manufacturers but also reflects the explosive growth trend of the entire market demand.

Similar signals have also been released on Alibaba Cloud's main stage.

At the World Artificial Intelligence Conference, Zhou Jingren, CTO of Alibaba Cloud, announced the latest progress of Tongyi Large Model and Alibaba Cloud's Bailian Platform. In the past two months, the download volume of Tongyi Qianwen's open-source model has doubled, exceeding 20 million downloads. The number of Alibaba Cloud Bailian's service customers has increased from 90,000 to 230,000, with a growth rate of over 150%.

When it comes to large models, compared to comparing parameters, domestic leaders now seem more willing to tell the market how well their large models work, how many people use them, and how they can be used next, among other things related to the implementation of scenario applications.

And investors represented by Zhu Xiaohu have also begun to look for investment opportunities in large models at the application level.

The market trend has changed, and it's not just the leaders' tone that has shifted.

Where are the "super capable" applications?

"In the AI era, 'super capable' applications are more important than 'super apps' that only focus on DAU." At the World Artificial Intelligence Conference, Robin Li tried to draw a conclusion about the next trend in large model application development.

However, understanding "super capable" applications may not be difficult, but the market's outstanding question is how such applications are developed and how they are promoted to the general public.

Based on the current industry performance, Smart Relativity believes that there are several considerations worth exploring.

Firstly, behind "super capable" applications, the iteration and adaptation of large model technology are necessary.

Industry trends mostly converge on the same path. The MoE architecture iteration trend that has emerged in the large model field this year represents that large models are technically supporting "rolling up" scenario applications.

Today, large models at home and abroad, such as OpenAI's GPT-4, Google's Gemini, Mistral AI's Mistral, xAI's Grok-1, Kunlun Wanwei's Tiangong AI, Inspur Information's Yuan 2.0-M32, and Tongyi Qianwen Team's Qwen1.5-MoE-A2.7B, have all adopted the MoE architecture.

The MoE architecture achieves model sparsity and modularity by introducing expert networks and gating mechanisms, providing considerable feedback in areas such as data processing, computational resource allocation, and output result optimization. This provides crucial technical support for the implementation and promotion of large model scenario applications.

For example, Microsoft has proposed an end-to-end MoE training and inference solution called DeepSpeed-MoE, which reduces communication overhead by deeply optimizing MoE's communication in parallel training, achieving efficient model parallelism. In addition, DeepSpeed-MoE also proposes an expert ranking mechanism based on fine-tuning, which can dynamically adjust the allocation of input samples to experts based on expert losses during training, improving performance.

Secondly, "super capable" applications imply a more commercialized ecological competition.

While the technology is sound, unclear commercialization paths can still lead to market collapse today. Recently, Microsoft's official website updated a notification announcing that "GPT Builder will be discontinued." The GPTs that once sparked countless heated discussions and hype in the AI circle seem to be heading towards defeat.

Who can recall that the launch event where the GPTs concept emerged was described by the outside world as "OpenAI's iPhone moment."

OpenAI originally intended to leverage low-barrier technological capabilities and global developers to create a batch of "super capable" applications, but due to experience flaws caused by technical issues and unclear monetization policies, the commercialization path of the GPTs concept has always been obstructed, ultimately leading to its demise.

Most "super capable" applications are built on mature business ecosystems, and perhaps AI vendors worldwide need to recognize this point. Worth mentioning is that on the other side of the ocean, Alibaba Cloud led the open-source AI model community, ModelScope, which recently won the 2024 SAIL Star Award in November 2022.

After more than a year of development, the ModelScope community has become the largest and most active AI model community in China, gathering over 5,500 high-quality models and thousands of datasets, providing model and free computing power services to over 5.6 million developers. Perhaps, the ecological path that OpenAI failed to navigate will find new vitality in China.

Thirdly, "super capable" applications must germinate in industry scenarios.

Zhu Xiaohu's advice to large model entrepreneurs is, "Don't blindly believe in AI, focus on sharp scenarios and implement them as soon as possible." Scenarios are the cradle for incubating "super capable" applications, but looking deeper, we cannot only focus on scenarios; ultimately, we must also consider user feedback and value presentation.

Industries such as healthcare, education, finance, manufacturing, transportation, and agriculture are "high-incidence" scenarios for large model applications, but how well the intelligent agents or solutions created actually perform is ultimately up to the users themselves.

To B projects focus on efficiency. In the express delivery industry, using large models to help process orders can achieve "sending a package with one image and one sentence," eliminating the need for other cumbersome processes, reducing the time from over 3 minutes to 19 seconds. Moreover, more than 90% of after-sales issues are also resolved by large models. Such efficiency improvements can truly be called "super capable".

To C scenarios focus on users. Previously, during peak periods, Baidu's intelligent college entrance examination system had to answer over two million questions from examinees every day. For the 10 million examinees nationwide, this ratio is quite high. Such a user base can also be considered "super capable".

Today, large model applications cover various general and vertical scenarios such as text generation, data processing, PPT creation, marketing, customer service, and medical diagnosis. In fact, the market is not lacking in scenarios but rather lacks capable and effective applications. "Rolling up" applications requires finding users and value in scenarios.

*All images in this article are sourced from the internet.